[Kamilbenkirane]

BytePlus image models warn about output_format and background (UnsupportedParameterWarning) and drop them. Celeste has the unified ImageParameter.OUTPUT_FORMAT and ImageParameter.BACKGROUND, and OpenAI maps both, but BytePlus has no mapper for either. #520 left this as follow-up work ("That needs a BytePlus OutputFormatMapper").

The ModelArk Image generation API documents both request fields:

• output_format: "string Default jpeg | Image format". "Specifies the file format of the generated image. Valid values: png, jpeg". Supported models: "Seedream 5.0 pro, Seedream 5.0 flash, Seedream 5.0 lite".
• background: "string Default opaque | Image alpha channel". "Controls whether to generate an image with an alpha channel. Valid values: transparent [...] opaque". Supported models: "Seedream 5.0 pro, Seedream 5.0 flash". Transparent mode "is supported only for image-to-image generation with exactly one input image that has an alpha channel", and "If output_format is set to jpeg, the request returns an error."

Fix:

• src/celeste/providers/byteplus/images/parameters.py: wire OutputFormatMapper (output_format) and BackgroundMapper (background).
• src/celeste/modalities/images/providers/byteplus/parameters.py: the two mappers are bound to ImageParameter.OUTPUT_FORMAT / ImageParameter.BACKGROUND and added to BYTEPLUS_PARAMETER_MAPPERS.
• src/celeste/modalities/images/providers/byteplus/models.py, with literals verbatim from the reference:
  • OUTPUT_FORMAT: Choice(["png", "jpeg"]) on seedream-5-0-260128 (lite), dola-seedream-5-0-pro-260628 and dola-seedream-5-0-flash-260915;
  • BACKGROUND: Choice(["transparent", "opaque"]) on pro and flash.
  • Seedream 4.0 / 4.5 get no constraint, so a value passes through unvalidated, by design.
  • The transparent-mode limits (image-to-image only, one alpha input, no jpeg) are mode-conditional, so they stay with the provider.
• Output MIME is unchanged: the artifact is still labeled from the response's declared output_format (fix(byteplus): label image outputs with their declared format instead of PNG #520).
• tests/unit_tests/test_parameter_wire_contracts.py: output_format and background rows for the BytePlus images list, which also joins the none-omits list.
• CHANGELOG.md: one Features line under [Unreleased].

Validation:

• make ci passes (766 tests).
• Request body for dola-seedream-5-0-flash-260915 with output_format="jpeg", background="transparent":
  • origin/main: neither field is sent, and two UnsupportedParameterWarnings are raised;
  • this branch: output_format: "jpeg" and background: "transparent" are sent, with no warnings.
• Live, through celeste's own client:

Call	artifact.mime_type	Downloaded bytes
flash generate(output_format="jpeg", quality="1K")	image/jpeg	ff d8 ff e0 (JPEG), 1248×832, 182,610 B
flash edit(image=<512×512 RGBA PNG, red disc on transparent>, background="transparent", quality="1K")	image/png	PNG, 1024×1024, rgba (IHDR color type 6). Alpha: 89.7% of pixels below 32 (background), 9.8% at 224 or above (the blue star)
lite generate(output_format="png", quality="2K")	None	PNG, 2048×2048, rgb24 (BytePlus defaults to jpeg, so the field was applied)

• Lite applies output_format but does not declare it in the response. The reference lists data.output_format only for 5.0 pro and flash, so lite output stays unlabeled. This is the same as fix(byteplus): label image outputs with their declared format instead of PNG #520: an undeclared format stays None, as for OpenAI, BFL and Topaz Labs. Streaming events document no output_format either.